摘要
Abstract
To address the automated detection needs in egg quality assessment and laying hen rearing condition monitoring,this paper proposes the YOLOv8-CTAC model to solve the problem of automatic segmentation of rough calcified matter on the surface of sand-shelled eggshells.Aiming at the shortcomings of YOLOv8s-Seg model in handling multi-scale information,feature expression,and attention to key regions,this study optimizes the feature extraction and expression ability of the model by integrating the Scale Sequence Feature Fusion(SSFF)module,Triple Feature Encoding(TFE)module,and Channel and Position Attention Mechanism(CPAM)module.Meanwhile,to cope with the sand shell category imbalance problem,the Varifocal Loss(VFL)function is introduced.The experimental results show that the YOLOv8-CTAC model exhibits significant improvements in both bounding box and mask evaluation aspects.Compared to the YOLOv8s-Seg model,it achieves enhancements of 6.7%in accuracy,8.3%in recall,and 7.4%in mean Average Precision(mAP)for bounding box evaluation.For mask evaluation,it improves by 8.3%in accuracy,8.9%in recall,and 8.2%in mAP.Moreover,compared to mainstream algorithms such as Mask R-CNN,SOLOv2,YOLOv8n-Seg,and YOLOv8s-Seg,it achieves an average improvement of 3.2%,10.1%,10.3%,and 6.7%respectively in mAP,significantly enhancing the detection performance in complex sandy-shell regions.This provides robust technical support and methodological assurance for the automated detection and segmentation tasks of sandy-shell eggs.关键词
深度学习/多尺度特征提取/图像分割/YOLOv8s/沙壳蛋Key words
Deep learning/multi-scale feature extraction/image segmentation/YOLOv8s/sand-shelled eggs分类
信息技术与安全科学
